Re: iSCSI: SCSI Iniitator port: a simple approach, acceptable?

[Santosh Rao <santoshr@cup.hp.com>]

Somesh Gupta wrote:

> > The initiator task tag & cmdsn can be passed down from the session
> > manager to the HBA in a mailbox command, as a part of issuing a new
> > iscsi command, and the (exp_cmdsn, max_cmdsn) values can be updated to
> > the host resident session manager by the HBA as a part of conveying a
> > mailbox command completion. There are no extra host interrupts in normal
> > I/O paths due to the presence of a host resident session manager.
> >
> > If you're referring to locking overheads, then, these can be optimized
> > out.
> 
>   You sort of answered the question yourself. As a cache line (containing
>   common numbers) ping-pongs
>   between processors that handle interrupts for different adapters, its
>   adds significant overhead. A 2GHz processor in a CC-Numa box sharing
>   cache lines between multiple processors?

I think the above can be optimized out by some proper design. I don't
think locking or cache thrashing are significant performance inhibitors
that cannot be worked around in this case. 


> 
>   This does not include the cost/complexity of managing multiple queues,
>   (and additional cache lines to share)

What multiple queues ? (I don't see any major issues here either.)

>   and the fact that the "rare" retransmit has the potential to slow
>   everything down to the speed of the slowest connection.

But, this is not a drawback of multi-connection session. It is a
drawback due to the fact that the iscsi protocol chose to impose strict
command ordering across the session [,which is not really required. But,
I won't start that thread again !].


> >
> >
> >
> > Somesh Gupta wrote:
> > >
> > > Jim,
> > >
> > > Multiple connections, especially when spread across multiple
> > > adapters will give performance that will be much less than
> > > that of multiple sessions in parallel (of course when there
> > > are multiple connections/session over multiple adapters - the
> > > adapters have to be able to take the session id from the
> > > host software - I obviously don't get the point of the debate).
> > >
> > > There is no way a common synchronization point for
> > > command sequencing and flow control at the host level will not
> > > cause sub-standard performance. I have argued this point
> > > many times before, only to have it be brushed aside. Of course,
> > > only time will tell.
> > >
> > > Somesh
> > >



-- 
##################################
Santosh Rao
Software Design Engineer,
HP-UX iSCSI Driver Team,
Hewlett Packard, Cupertino.
email : santoshr@cup.hp.com
Phone : 408-447-3751
##################################